junolee / joke-recommender

Built recommendation system based on a dataset of over 4 million ratings of jokes using GraphLab

Joke Recommender System

Recommendation system built based on a dataset of over 4 million ratings of jokes. Implemented an item similarity recommender and factorization recommenders using GraphLab.

• Task: Build a recommendation system for jokes
• Data: User ratings ranging from -10 to 10
• Scoring: Mean rating for top 5% of jokes predicted by recommender
• Tools: GraphLab, Pandas

Exploratory Data Analysis

Distribution of Ratings, Average Joke Rating, Number of Ratings per Joke:

Choosing a Model

• If you care about accurately predicting the rating a user would give a specific item - Factorization recommender
• If you care about ranking performance, instead of simply predicting the rating accurately - Item Similarity Recommender or Ranking Factorization recommender

More details here.

Tuning Parameters for Factorization Recommenders

Number of Latent Features:

• Factorization Recommender:
  • Default: 8
  • Optimal: 2
• Ranking Factorization Recommender:
  • Default: 32
  • Optimal: 8

Regularization: Default

Model Results

• Factorization Recommender (num_factors=2)
  • Score: 2.5001
• Ranking Factorization Recommender (num_factors=8)
  • Score: 2.1726
• Item Similarity Recommender (similarity_type=pearson)
  • Score: 2.5301
• Item Similarity Recommender (similarity_type=cosine)
  • Score: 2.3010
• Item Similarity Recommender (similarity_type=jaccard)
  • Score: 1.5739